use cast op for gpu

chraac · Feb 7, 2025 · 05df736 · 05df736
1 parent c667db2
commit 05df736
Show file tree

Hide file tree

Showing 2 changed files with 6 additions and 7 deletions.
diff --git a/ggml/src/ggml-qnn/backend-ops.cpp b/ggml/src/ggml-qnn/backend-ops.cpp
@@ -448,7 +448,6 @@ bool ggml_qnn_supports_matmul_op(ggml_backend_qnn_device_context *ctx, const ggm
             }
             // fall through, from test here, the convert op is super slow on NPU:
             //   https://github.com/usefulsensors/qc_npu_benchmark
-        case QNN_BACKEND_GPU:
             if (src0->type != src1->type || src0->type != op->type) {
                 // there's no convert op for GPU.
                 QNN_LOG_DEBUG(
@@ -457,6 +456,7 @@ bool ggml_qnn_supports_matmul_op(ggml_backend_qnn_device_context *ctx, const ggm
                 return false;
             }
             break;
+        case QNN_BACKEND_GPU:
         default:
             break;
     }

diff --git a/ggml/src/ggml-qnn/op-config-impl.cpp b/ggml/src/ggml-qnn/op-config-impl.cpp
@@ -307,10 +307,9 @@ qnn_tensor_ptr_t ggml_qnn_matmul_op_config::create_gather_nodes(QNNBackend devic
 bool ggml_qnn_matmul_op_config::create_convert_nodes(QNNBackend device, Qnn_GraphHandle_t graph_handle, const int rank,
                                                      qnn_tensor_array_t &tensor_inputs,
                                                      qnn_tensor_array_t &tensor_outputs) {
-    if (device == QNN_BACKEND_GPU) {
-        // there's no convert op for GPU, so we should create matmul nodes directly.
-        return true;
-    }
+    // there's no convert op for GPU, so we use cast instead
+    // https://docs.qualcomm.com/bundle/publicresource/topics/80-63442-50/MasterOpDef.html#cast
+    const std::string convert_op_name = (device == QNN_BACKEND_GPU) ? QNN_OP_CAST : QNN_OP_CONVERT;
 
     // create tensors for convert node
     auto tensor_type = get_tensor_type(tensor_inputs);
@@ -328,7 +327,7 @@ bool ggml_qnn_matmul_op_config::create_convert_nodes(QNNBackend device, Qnn_Grap
                                                              convert_in->get_dimensions(), tensor_type, rank, device,
                                                              graph_handle, _qnn_instance);
         auto convert = std::make_shared<ggml_qnn_single_op_config>(convert_name, QNN_OP_PACKAGE_NAME_QTI_AISW,
-                                                                   QNN_OP_CONVERT, _qnn_instance);
+                                                                   convert_op_name, _qnn_instance);
         convert->set_input_tensors({convert_in});
         convert->set_output_tensors({convert_out});
         tensor_inputs[i] = convert_out;
@@ -343,7 +342,7 @@ bool ggml_qnn_matmul_op_config::create_convert_nodes(QNNBackend device, Qnn_Grap
                                                             convert_out->get_dimensions(), tensor_type, rank, device,
                                                             graph_handle, _qnn_instance);
         auto output_convert = std::make_shared<ggml_qnn_single_op_config>(convert_name, QNN_OP_PACKAGE_NAME_QTI_AISW,
-                                                                          QNN_OP_CONVERT, _qnn_instance);
+                                                                          convert_op_name, _qnn_instance);
         output_convert->set_input_tensors({convert_in});
         output_convert->set_output_tensors({convert_out});
         tensor_outputs.front() = convert_in;